Looking at `bytes-allocated` usually works, but sometimes lets
memory use spiral out of control. Looking at `current-memory-bytes`
is more reliable, but still makes peak memoy use too high.
Combining those values doesn't limit the peak well enough.
Try the more obvious (in retrospect) approach of comparing
`bytes-allocated` changes and `current-memory-bytes` changes
separately, and triggering a GC if either grows enough.
Commit 1811193285 caused Racket CS to have much higher peak memory
use. Adjust the heuristic again to trigger a major GC when the
`current-memory-bytes` value is the post-GC `bytes-allocated` plus the
post-GC `current-memory-bytes`, which means waiting until
another `bytes-allocated` bytes are allocated (instead of waiting
until the number of newly allocated bytes also catches up to overhead,
such as unserused pages due to locked objects).
Implement a parameter as a Chez Scheme wrapper procedure,
instead of an applicable record. A wrapper procedure can be
applied more directly, saving 10-20% of the time for some
parameter lookups.
The grammar for pregexps includes:
| \p{‹property›} Match (UTF-8 encoded) in ‹property›
| \P{‹property›} Match (UTF-8 encoded) not in ‹property›
and <property> is defined as:
‹property› ::= ‹category› Includes all characters in ‹category›
| ^‹category› Includes all characters not in ‹category›
That is to say, there are two independent ways to negate one of
these character classes. The Racket implementation of regexps
(as opposed to the C implementation) does not recognize negated
categories. This PR fixes that.
Weak pairs, ephemeron pairs, some symbols, and some ports were handled
incorerctly when locked multiple times.
original commit: 847fc1c84496f67cd363c8411d0023339f4d6246
The `unlock-object` operation was O(N) with N currently locked objects
--- so, O(N^2) to lock N objects and then unlock them --- because
locked objects were stored in and searched in a global list. Also, GC
was O(N) at any generation with N locked objects across generations,
since every locked object was scanned.
Fix these poblems so that locking and unlocking is practically O(1)
and GC is not poportional to locked objects. More precisely, locking
and unlocking is now O(C) for locking an individual object C times to
be balanced by C unlocks. (Since multiple locks on a single object
is rare, this performance seems good enough.)
The implementation replaces the global list with segment-specific
lists. Backpointers are managed using the general generational
support, so that unmodified, old-generation locked objects do not
need to be swept duing a new-generation collection.
original commit: a57d256ca73a3d507792c471facb7e35afbe88b3
Some parts of the GC meant to traverse all objects on a page of
'atomic-interior or 'interior objects used "<" to detect the end of
the page, but "<=" was needed. As a result, things could sometimes go
wrong for the last object on a page for platform and size combinations
where the last object ends exactly at the end of the page.
This change consistenly computes the iteration end in a way that makes
both "<" and "<=" work.
Using MPFR bindings from the `math` library could trigger a problem
(but it's difficult to provoke the problem in a reasonably small
example --- difficult enough that I couldn't do it).
Use `current-memory-bytes` instead of `bytes-allocated` to determine
major GCs, because the latter doesn't include enough (perhaps missing
finalized values). For example, the repair avoids unbounded memory use
from
(let loop ([i 0])
(malloc 6400 'atomic-interior)
(loop))
due to finalizers that pile up faster than they are run.
When reading just the vector section of the Reference, it's
not obvious that `immutable?` distinguishes immutable vectors,
so add a note there. And the same for sttrings, etc. The section
on boxes was missing the usual paragraph on mutability.
Based on a suggestion from Shriram.
When a linklet is too large to pass to Chez Scheme whole, then
names for the procedures that are individually compiled need to
be extracted from 'inferred-name for reference in the wrapper.
Closes#2787
Enables native dark mode UI elements in macOS 10.14 and above. Adds the
'NSRequiresAquaSystemAppearance' key with a value of 'false' to the
Info.plist file, while allows UI elements to match the system theme even
when not building against the latest SDK.
When opening the write end of a fifo that doesn't have a reader
already, the old implementation could allow writing bytes that are
discarded. This new implementation uses a blocking `open` in a
`pthread`, and that way the write routines know whether the stream is
ready for writing or not.
The difference is visible in the Racket API in a two places:
`subprocess` needs to wait until a fifo writer is connected before
attempting to dup the corresponding file descriotor, and more
generally a use of `unsafe-port->file-descriptor` needs to wait. The
former blocking operation is now build into `subprocess` (and
documented), but the burden is place on callers of
`unsafe-port->file-descriptor` to wait is necessary.
The new `port-waiting-peer?` predicate exposes the waiting state,
while `sync` is sufficient to wait for a peer.
Closes#2784
On platforms other than Windows and MacOS, locale encoding (inclduing
path <-> string conversion) opened a converter for each separate
operation. That can be slow on some OSes, so cache converters used for
locale conversions.
Relevant to #2781
Allowing #f as an allocator avoids problems composing `allocator` with
foreign-function lookup where failure is anticipated and implemented
as #f. For example, `g_settings_new` in the "gui-lib" package's
"mred/private/wx/gtk/gsettings.rkt" can be #f if the libgio libray is
too old, in which case there won't be an attempt to use
`g_settings_new`.
In some cases, (vector x 2 3 3 3) was pretty-printed as "(vector x 2 3)" when
print-vector-length was enabled.
Also print "(fxvector)" instead of "(fxvector )".
